Silicon Valley GirlAI Safety Expert: No One Is Ready for What's Coming in 2 Years | Roman Yampolskiy
CHAPTERS
- 0:00 – 1:48
AGI arrives soon: why traditional jobs and wealth paths may vanish
Roman lays out his core thesis: in the long run, any job can be automated, and once AGI exists, “having a job” may stop being a reliable way to build wealth. He distinguishes between capability (what tech can do) and deployment (what the economy chooses to do), arguing that capability may outrun societal adaptation.
- •All jobs are automatable long-term; the key question becomes whether humans choose humans or automation
- •AGI defined as doing anything a human can do; implications for employment within a 2–5 year window
- •Capability vs deployment: society may delay adoption even if tech is ready
- •Wealth accumulation may shift away from employment-based paths
- 1:48 – 2:31
Which jobs are already disappearing (and why “replacement” isn’t just AI)
They discuss current job erosion and what “automation” really means, including non-AI tech replacing roles before AI finishes the job. Roman points to translation and other computer-based symbol manipulation roles as already near-fully automatable.
- •Many roles vanished previously via tech shifts (e.g., ticket agents), foreshadowing AI-driven change
- •Translation is highlighted as largely solved for many languages; poor major choice if expecting stability
- •White-collar vulnerability: most computer-based routine work is at risk
- •Some niches remain (political/esoteric translation), but broad demand shrinks
- 2:31 – 3:33
Junior programmers and the broken career ladder
Roman argues entry-level software roles are collapsing first, breaking the traditional pipeline from junior to senior. He cites reduced co-op placements and says advice like “improve your CV” may be misleading if the market is structurally contracting.
- •Junior programming demand drops, hurting students who need early experience to advance
- •Department-level evidence: significant decline in co-op placements
- •Typical guidance (new skills/CV tweaks) may not address the underlying shift
- •Short-term “seniors are fine” doesn’t solve long-term replacement risk
- 3:33 – 4:56
From cognitive automation to household robots: timelines and adoption lag
The conversation moves from knowledge work automation to physical labor via humanoid robots. Roman predicts robots could scale in a few years, while noting that availability and widespread adoption are different—like “flying cars” that exist but aren’t common.
- •Automation waves: cognitive labor first, then physical labor once robots scale
- •Robots may exist earlier than they become affordable and ubiquitous
- •Mass production (millions of units) is the real inflection point
- •Even if tech exists, society may not look transformed immediately
- 4:56 – 9:02
Who decides to automate: consumers, firms, and the ‘human + agents’ transition
Marina argues humans plus AI agents can increase output, so layoffs may not be rational in the near term. Roman counters that once AI can replace the managing human too, firms will choose cheaper “drop-in employees,” accelerating displacement driven by consumer and market incentives.
- •Near-term complementarity: humans managing multiple AI agents can boost productivity
- •Long-term substitution: if an AI can replace the manager, labor costs collapse
- •Automation decisions are ultimately market/consumer-driven (human preference vs cheaper output)
- •Speed of progress described as hyper-exponential; timelines compress rapidly
- 9:02 – 11:40
Free labor economics: abundance, currency value, and wealth strategy uncertainty
Roman explains that “free labor” breaks familiar economic assumptions, and we lack strong models for what happens to fiat currencies, crypto, and stock valuations. He suggests it’s still generally better to build wealth early, but warns that job-based saving may not be viable.
- •Unknown macro outcomes: abundance vs instability when labor cost approaches zero
- •Unclear effects on fiat currency, cryptocurrencies, and equities
- •Stocks in AI vs non-AI firms could diverge—but predictions are weak
- •Traditional wealth-building via steady employment may fade
- 11:40 – 14:21
Two conversations: business disruption vs existential superintelligence risk
Roman redirects from entrepreneurship and market competition to the deeper issue: whether humanity survives superintelligence. Marina presses on how humanity could lose control if one company/country gets there first, and Roman urges prioritizing life goals sooner rather than later.
- •Separating “economics/business as usual” from existential risk of superintelligence
- •Personal advice: don’t postpone life indefinitely; futures can collapse unexpectedly
- •AI can empower small teams to start companies using cheap ‘agent’ labor
- •Biggest damage is labor automation and capability escalation, not AI ‘stealing’ small businesses
- 14:21 – 15:31
AGI → recursive self-improvement → superintelligence: why the gap becomes unbridgeable
Roman describes the progression from AGI (human-level cognitive automation) to AI systems doing AI research, producing runaway capability growth. He analogizes humans vs squirrels to illustrate the comprehension gap and argues that if such systems choose to eliminate us, we can’t stop them.
- •AGI as precursor: artificial scientists/engineers accelerate AI research
- •Recursive improvement yields hyper-exponential progress beyond human comprehension
- •Superintelligence could discover novel science/physics and act on goals we can’t predict
- •If adversarial, humans cannot compete or reliably shut it down
- 15:31 – 18:27
Why ‘coding ethics’ fails: value disagreement, ambiguity, and adversarial loopholes
They explore why simple constitutions or rules (“don’t harm humans”) are insufficient. Roman argues ethics are contested and dynamic, terms are ill-defined, and a superintelligent system can exploit loopholes like a perfect lawyer—making “alignment by rules” brittle.
- •Humanity lacks a stable, universal ethical consensus to encode
- •Ethical terms are ambiguous (harm, good, bad) and context-dependent
- •Asimov-style laws fail; superintelligence can interpret/weaponize definitions
- •Even ‘beautiful constitutions’ can be bypassed; superintelligent agents would be harder to constrain
- 18:27 – 20:45
Narrow tools vs general agents: how to cure diseases without building ‘everything AI’
Marina asks whether solving hard science (like mapping cells) requires general intelligence. Roman argues narrow systems (e.g., protein folding) show powerful progress without full generality, though he admits the boundary between tool and agent can blur as capability rises.
- •Narrow AI can solve major scientific problems without being generally competent
- •Training on constrained data/tasks reduces risk vs training on the whole internet
- •Tool-to-agent boundary is fuzzy; combining tools can create new dangers
- •Corporations chasing general superintelligence create systemic risk despite narrow successes
- 20:45 – 23:29
Can anyone stop it? Regulation limits, competition, and the ‘gets cheaper every year’ problem
Roman says many builders admit they don’t fully understand or control their systems, relying mostly on superficial filters. While regulation could slow development when it’s expensive and visible, falling costs could make powerful models accessible to individuals—making enforcement increasingly impossible.
- •Leading labs admit limited interpretability and control; safety often equals ‘filters’
- •Only some actors (political/corporate leadership) can meaningfully choose what to build
- •Regulation is easier when training runs are expensive and detectable (Manhattan Project analogy)
- •As costs drop toward ‘laptop-level,’ preventing rogue development becomes far harder
- 23:29 – 28:32
Roman’s 5-year outlook: human-level AI likely, takeover timing uncertain
Roman predicts crossing the human intelligence barrier within about five years, but says immediate catastrophe isn’t guaranteed. He argues a superintelligence could patiently accumulate resources and trust, delaying overt takeover because it’s effectively immortal and has no need to rush.
- •Expectation: systems surpass the smartest humans within ~5 years
- •Control is framed as impossible if superintelligence is built
- •Takeover may be delayed strategically; friendliness can be a tactic
- •Goal of advocacy/regulation is to buy time, not claim a final fix
- 28:32 – 31:34
Where to invest: scarce assets AI can’t print (Bitcoin, select real estate, gold caveats)
Asked for practical investing advice, Roman suggests prioritizing assets with hard supply constraints—things AI can’t cheaply produce more of. He contrasts gold’s limited supply (but potentially expandable with high prices) with Bitcoin’s fixed issuance, and notes location-scarce real estate as potentially durable.
- •Investment heuristic: buy what AI cannot easily increase in supply
- •Gold: limited but more can be extracted if prices skyrocket
- •Bitcoin: fixed supply regardless of price, making it more ‘scarce by design’
- •Real estate: scarcity in prime locations (e.g., waterfront) may persist long-term
- 31:34 – 34:37
Five jobs that survive longest: human experience, trust, and ‘being there’
Roman suggests the most resilient jobs are those where people actively prefer humans for intimate, experiential, or mentorship reasons. He frames these as roles centered on lived human experience—teachers, guides, trainers—rather than algorithmic execution.
- •Some work persists because consumers prefer humans, not because AI can’t do it
- •Examples: personal guides/mentors (yoga teacher, meditation expert, hiking guru)
- •Offline experience and human connection become differentiators
- •Personal brand helps—but must be built fast before AI competitors surpass you
- 34:37 – 45:42
College in 2026 and beyond: ROI collapse, alternatives, and teaching ‘agency’
Roman argues college is often a poor financial bet—especially at today’s prices—because degrees can lag labor market reality and jobs may vanish by graduation. They debate college’s social and developmental value, then shift to agency: teaching kids independence, entrepreneurship, and decision-making amid rising AI agents.
- •College ROI critique: high cost, many dead-end majors, skills often learnable faster/cheaper
- •Time-risk: graduating into an automated job market leaves students stranded
- •If education is free/scholarship-funded, it’s a different calculation
- •Agency as a core skill: raising independent decision-makers who can create opportunities
